-
Notifications
You must be signed in to change notification settings - Fork 11
/
Copy pathdataset.py
195 lines (173 loc) · 8.16 KB
/
dataset.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
import os.path as osp
import copy
import json
import math
import random
import torch
import numpy as np
from torch_geometric.data import Data
from blocks.conv import cal_edge_attr
from tools.cut_stroke import cutoff
from tools.dislocation import dislocate
from tools.meaningless_stroke import additional_label
from tools.center_sketch import rotate_theta, add_normal_noise
class SketchData(Data):
def __init__(self, stroke_idx=None, x=None, edge_index=None, edge_attr=None, y=None,
pos=None, norm=None, face=None, **kwargs):
super(SketchData, self).__init__(x, edge_index, edge_attr, y, pos, norm, face, **kwargs)
self.stroke_idx = stroke_idx
self.stroke_num = max(stroke_idx) + 1
def __inc__(self, key, value):
if 'index' in key or 'face' in key:
return self.num_nodes
elif 'stroke' in key:
return self.stroke_num
else:
return 0
class SketchDataset(torch.utils.data.Dataset):
def __init__(self, opt, root, class_name, split='train', permutation=False):
self.class_name = class_name
self.split = split
self.pt_dir = osp.join(root, '{}_{}.pt'.format(self.class_name, self.split))
self.json_dir = osp.join(root, '{}_{}.ndjson'.format(self.class_name, self.split))
self.permutation = permutation
self.perm_args = opt.perm_arg
self.perm_type = opt.perm_type
self.out_segment = opt.out_segment
if self.permutation:
if opt.is_train:
if self.perm_type == 'dislocation':
self.processed_data = self._process()
self.augmentation(perm_args=[0.1, 0.3]) # dislocation
elif self.perm_type == 'cutoff':
self.processed_data = self._process()
self.augmentation(perm_args=[0.8, 3.2]) # cut off
elif self.perm_type[:8] == 'meanless':
self.processed_data = self._process(perm_arg=0)
self.augmentation(perm_args=[0.15, 0.2]) # meanless
else:
raise NotImplementedError('Permutation Type {} is not implemented!'.format(self.perm_type))
else:
self.processed_data = self._process(opt.perm_arg)
else:
if osp.exists(self.pt_dir):
self.processed_data = torch.load(self.pt_dir)
else:
self.processed_data = self._process()
def __getitem__(self, index):
return self.processed_data[index]
# if self.permutation:
# # dislocation
# # sketch = self.processed_data[index]
# sketch = copy.deepcopy(self.processed_data[index])
# dislocate_stroke_num = int(sketch.stroke_num*0.5)
# idxs = np.random.choice(sketch.stroke_num, dislocate_stroke_num, replace=False)
# for idx in idxs:
# _idx = (sketch.stroke_idx == idx).int().reshape(-1, 1)
# offset = self.perm_args*(torch.rand(2) - 0.5)
# offset = offset.repeat(256, 1) * _idx
# sketch.x += offset
# # # noise
# # sketch = copy.deepcopy(self.processed_data[index])
# # # _sketch = self.processed_data[index]
# # sketch.x = sketch.x*255 + torch.randn(sketch.x.shape)*self.perm_args
# # # sketch_x = _sketch.x + torch.randn(_sketch.x.shape)*self.perm_args
# # # rotation
# # sketch = self.processed_data[index]
# # theta = random.randint(0, self.perm_args) * math.pi/180
# # if random.randint(0, 1):
# # theta = -theta
# # rotate_matrix = torch.tensor([[math.cos(theta), math.sin(theta)],[-math.sin(theta), math.cos(theta)]])
# # sketch.x = torch.matmul(sketch.x, rotate_matrix.to(sketch.x.dtype).to(sketch.x.device))
# # norm
# max_point = torch.max(sketch.x, dim=0).values
# min_point = torch.min(sketch.x, dim=0).values
# sketch.x = (sketch.x - min_point) / (max_point - min_point)
# return sketch
# else:
# return self.processed_data[index]
def __len__(self):
return len(self.processed_data)
def augmentation(self, perm_args):
for perm_arg in perm_args:
self.processed_data.extend(self._process(perm_arg))
def _process(self, perm_arg=None):
if perm_arg is not None:
print('Processing with augment param {} ...'.format(perm_arg))
else:
print('Processing without augment param ...')
raw_data = []
with open(self.json_dir, 'r') as f:
for line in f:
raw_data.append(json.loads(line)["drawing"])
processed_data = []
for sketch in raw_data:
if perm_arg is not None:
if self.perm_type == 'rotate':
theta = perm_arg * math.pi/180
sketch = rotate_theta(sketch, theta)
elif self.perm_type == 'noise':
sketch = add_normal_noise(sketch, scale=perm_arg)
elif self.perm_type == 'dislocation':
sketch = dislocate(sketch, 1, perm_arg)
elif self.perm_type == 'cutoff':
sketch = cutoff(sketch, perm_arg)
elif self.perm_type == 'meanless_addition':
sketch = additional_label(sketch, perm_arg, label_num=self.out_segment-1, labeling_type='addition')
elif self.perm_type == 'meanless_random':
sketch = additional_label(sketch, perm_arg, label_num=self.out_segment, labeling_type='random')
else:
raise NotImplementedError('Permutation Type {} is not implemented!'.format(self.perm_type))
sketchArray = [np.array(s) for s in sketch]
stroke_idx = np.concatenate([np.zeros(len(s[0])) + i for i, s in enumerate(sketchArray)])
point = np.concatenate([s.transpose()[:,:2] for s in sketchArray])
# normalize the data (N x 2)
point = point.astype(np.float)
max_point = np.max(point, axis=0)
min_point = np.min(point, axis=0)
point = (point - min_point) / (max_point - min_point)
# point /= 255
# label: c (N,)
label = np.concatenate([s[2] for s in sketchArray], axis=0) # (N, )
# edge_index
edge_index = []
s = 0
for stroke in sketchArray:
# edge_index.append([s,s])
for i in range(len(stroke[0])-1):
edge_index.append([s+i, s+i+1])
edge_index.append([s+i+1, s+i])
# edge_index.append([s,s+len(stroke[0])-1])
s += len(stroke[0])
edge_index = np.array(edge_index).transpose()
sketch_data = SketchData(x=torch.FloatTensor(point),
edge_index=torch.LongTensor(edge_index),
stroke_idx=torch.LongTensor(stroke_idx),
y=torch.LongTensor(label))
sketch_data.edge_attr = cal_edge_attr(sketch_data.edge_index, sketch_data.x)
processed_data.append(sketch_data)
# torch.save(processed_data, self.pt_dir)
return processed_data
if __name__ == "__main__":
import os
import torch
from options import TestOptions
from utils import load_data
opt = TestOptions().parse()
opt.perm_arg = 0.3
dataloader = load_data(opt, datasetType='test', permutation=True)
for e in range(3):
for i, batch in enumerate(dataloader):
print(batch.x)
# print(batch.stroke_idx)
# print(batch.stroke_idx[:300])
# print(batch)
# print(batch.num_graphs)
# print(batch.batch.shape)
# row, col = batch.edge_index
# offset = batch.pos[col] - batch.pos[row]
# print(offset)
# dist = torch.norm(offset, p=2, dim=-1).view(-1, 1)
# edge_attr = torch.cat([offset, dist], dim=-1)
# print(batch.edge_attr)
break